A real measurement · not simulated
Live benchmark
Everything the rest of the site calculates comes from a physical model. This is where it gets checked: this page runs the job for real against the card in the computer it is running on, shows the tokens as they come out and saves the measurement. If a number on this site is a lie, it shows up here.
—
- Model
- —
- Quantisation
- —
- File
- —
- KV cache
- —
- Context
- —
- Engine
- llama.cpp
A single slot: runs go one at a time.
This page only works on the computer that has the card
A live run needs to talk to the engine running on that card. From anywhere else there is nothing to measure: what you would see would be made-up numbers, and that is precisely what this page exists to prevent.
The measured figures the rest of the site uses were taken this way, and they are in the method.
1 · Pick the run
Each one is a real administrative job, with its own prompt.
2 · The card at work
3 · Every measurement
Every run is saved. The medians leave out the failed ones and the cold starts.
Every measurement in the history is of the same model: Qwen3.8 27B · quantised to 4 bits (Qwen3.8-27B-UD-Q4_K_XL.gguf, 17.55 GB) with the KV cache at q8_0 (8 bits).
| When | Run | In | Out | Reads (t/s) | Writes (t/s) | Total | Watts |
|---|---|---|---|---|---|---|---|
| Loading the history… | |||||||
How it is measured
The reads (prefill: digesting the document) and writes (decode: drafting the answer) figures are not timed by the browser: they come from llama.cpp itself in the last chunk of the stream. Power and memory are read with nvidia-smi in the middle of generation, which is when the card is genuinely working.
The first run after a while idle usually comes out much slower than normal: the card was asleep. Those runs are saved anyway, marked as a cold start, and they do not count towards the medians.